A scoring model can pass its math checks and still give a fragile decision once the weights, missing data, or operating queue move.
Our SSRN working paper ran 24,000 synthetic work-reasoning profiles through four software team topologies, then changed the inputs, removed domains, moved the weights, and pushed the delivery queue toward saturation. The useful result was not one polished score. It was a map of where the answer moved.
Small weight perturbations changed the highest-scoring topology for 10.7% of profiles on average, while one sampled draw reached 48.4%. At noise levels of 0.05 and 0.10, mean absolute score error reached 1.67 and 3.50 points while rank correlation fell, and removing one domain moved the error again.
The queue test made the operating risk plain: at 50% utilization, estimated wait was one service-time unit; at 95%, it was 19. A strong engineer plus a useful AI agent can still lose inside an overloaded review system.
Synthetic evidence does not validate a human construct or prove job performance, but the experiment shows a CTO what must be governed before a human study: weights, missing data, sensitivity bands, topology assumptions, queue pressure, fairness, accommodation, and the accountable human decision.
I turned the paper into a plain-English TeamStation field guide with the full results, limits, and next scientific gates. The reason to read it is the failure map, not a magic number:
https://teamstation.dev/research/articles/human-task-agent-alignment-stress-test
EngineeringScience #HumanAIAlignment #TeamTopologies #EngineeringTelemetry #TeamStationAI
Related TeamStation sources:
- Human Alignment in the Agentic AI Era
- Nearshore Engineering Team Models
- CTO Nearshore Strategy Control Center
- Free LATAM IT Talent Pricing Calculator
GitHub topic map:
Source asset:
https://teamstation.dev/research/articles/human-task-agent-alignment-stress-test
Top comments (0)